Diffusion Denoising for Low-Dose-CT Model

Runyi Li

TL;DR

This paper introduces Denoising Diffusion LDCT Model (DDLM), an unsupervised method for reconstructing high-quality CT images from low-dose scans. DDLM leverages pre-trained diffusion models and a novel conditional sampling strategy, eliminating the need for paired full-dose and low-dose data and offering comparable or superior performance to supervised methods with faster inference.

What Problem Does This Paper Address?

Medical imaging, particularly Computed Tomography (CT), is crucial for diagnosis. However, the radiation dose associated with CT scans poses a health risk. Low-dose CT (LDCT) aims to reduce this risk by using less radiation, but this often results in images with significant noise and blur, hindering accurate diagnosis. While deep learning methods have shown promise in improving LDCT image quality, they typically require large datasets of paired full-dose and low-dose images for supervised training. This is a significant bottleneck, as obtaining such paired data is difficult and expensive. Furthermore, supervised methods can be overly dependent on the specific imaging parameters of the training data, limiting their generalizability. This paper addresses the need for an effective LDCT reconstruction method that is unsupervised, generalizable, and doesn't require paired data.

Methods

The core of DDLM is built upon Denoising Diffusion Probabilistic Models (DDPMs), a class of generative models that have shown remarkable success in image synthesis. DDPMs work by gradually adding noise to an image until it becomes pure noise, and then learning to reverse this process to generate a clean image from noise.

The authors introduce several key innovations:

  • Conditional Sampling: Standard diffusion models generate unconditionally. DDLM implements conditional sampling by incorporating the low-dose CT image (y) into the denoising process. This is achieved through a range-null space decomposition of the measurement matrix, which helps to disentangle the signal from noise and constraints.
  • Unsupervised Approach: DDLM operates in an unsupervised manner. It utilizes a pre-trained diffusion model and requires no further training or fine-tuning. The low-dose CT image is directly used as input during the sampling (generation) phase.
  • Enhanced DDLM+: An enhanced version, DDLM+, is proposed to further improve performance. This version incorporates:
  • Time-travelling trick: This technique involves using future states in the reverse diffusion process to guide the current step, leading to more accurate and visually realistic results.
  • Low-pass filter: A low-pass filter is applied to extract domain-invariant features from the input LDCT image, enhancing generalization across different imaging devices and protocols.
  • Adaptive noise intensity consideration: The method adaptively determines the influence of the input image's spectral noise versus the current denoising state, aiming to avoid interference and improve generation quality.

Datasets

  • AAPM Challenge Dataset: This dataset was used for pre-training the diffusion models and for testing. It includes full-dose CT images (ground truth) and quarter-dose CT images.
  • Piglet Dataset: This dataset was used to evaluate the generalization ability of the proposed methods on out-of-domain data.

Key Insights and Ideas

  • Unsupervised LDCT is Achievable and Effective: The paper demonstrates that high-quality LDCT reconstruction can be achieved without the need for paired supervised data, overcoming a major practical hurdle.
  • Diffusion Models are Powerful for Inverse Problems: By conditioning diffusion models on the degraded input, their generative capabilities can be harnessed to solve challenging inverse problems like LDCT reconstruction.
  • Conditional Sampling via Range-Null Space Decomposition is Key: This technique allows the diffusion model to effectively "understand" and utilize the information from the low-dose CT image without explicit training on paired data.
  • Time-travelling and Domain-Invariant Features Boost Performance and Generalization: The DDLM+ enhancements highlight the importance of looking ahead in the diffusion process and filtering out domain-specific noise to achieve better visual quality and broader applicability.

Key Results

  • Comparable or Superior Performance: DDLM and DDLM+ achieve comparable or better performance (measured by PSNR, SSIM, and FID scores) than baseline methods, including supervised deep learning approaches like REDCNN and WGAN, on the AAPM dataset.
  • Faster Inference: DDLM offers significantly less inference time compared to some state-of-the-art methods, making it more practical for clinical use.
  • Improved Visual Quality: The DDLM+ version, with its enhancements, produces results with better visual realism.
  • Good Generalization: The proposed methods show promising results on out-of-domain datasets, indicating good generalization capabilities.

Limitations and Future Work

  • Recursive Reconstruction Pipeline: The paper acknowledges that the recursive nature of the diffusion process can limit efficiency. Future work will focus on improving speed, potentially through techniques like DPM-solver.
  • Broader Application: The authors plan to extend their work to other medical imaging domains and general image inverse problems.

Summary updated

Abstract

Low-dose Computed Tomography (LDCT) reconstruction is an important task in medical image analysis. Recent years have seen many deep learning based methods, proved to be effective in this area. However, these methods mostly follow a supervised architecture, which needs paired CT image of full dose and quarter dose, and the solution is highly dependent on specific measurements. In this work, we introduce Denoising Diffusion LDCT Model, dubbed as DDLM, generating noise-free CT image using conditioned sampling. DDLM uses pretrained model, and need no training nor tuning process, thus our proposal is in unsupervised manner. Experiments on LDCT images have shown comparable performance of DDLM using less inference time, surpassing other state-of-the-art methods, proving both accurate and efficient. Implementation code will be set to public soon.
BibTeX
@misc{li2023diffusion,
  author = {Runyi Li},
  title = {{Diffusion Denoising for Low-Dose-CT Model}},
  year = {2023},
  eprint = {2301.11482},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.11482},
}
Figure